Intro
PyTorch provides a variety of means to crop images. For example, torchvision.transforms provides several functions to crop PIL
images; PyTorch Forum provides an answer of how to crop image in a differentiable way (differentiable with respect to the image). However, sometimes we need a fully differentiable approach for the cropping action itself. How shall we implement that?
Theory: Affine transformation
Before reaching the answer, we need first to learn about the image coordinate system in PyTorch. It is a left-handed Cartesian system origined at the middle of an image. The coordinate has been normalized to range [ − 1 , 1 ] [-1,1] [−1,1], where ( − 1 , − 1 ) (-1,-1) (−1,−1) indicates the top-left corner, and ( 1 , 1 ) (1,1) (1,1) indicates the bottom-right corner, as pointed out by the doc.
Let ( x , y ) (x,y) (x,y) be the top-left corner of the cropped image with respect to the coordinate of the original image; likewise, we denote ( x ′ , y ′ ) (x',y') (x′,y′) as the bottom-right corner of the cropped image. It’s clear that ( x , y ) (x,y) (x,y) corresponds to ( − 1 , − 1 ) (-1,-1) (−1,−1) with respect to the cropped image coordinate system, and ( x ′ , y ′ ) (x',y') (x′,y′) corresponds to ( 1 , 1 ) (1,1) (1,1). We’d like a function f f f that maps from the cropped image system to the original image system for every point in the cropped image. Since only scaling and translation are involved, the function f f f can be parameterized by an affine transformation matrix Θ \Theta Θ such that
Θ = ( θ 11 0 θ 13 0 θ 22 θ 23 0 0 1 ) \Theta = \begin{pmatrix} \theta_{11} & 0 & \theta_{13}\\ 0 & \theta_{22} & \theta_{23}\\ 0 & 0 & 1\\ \end{pmatrix} Θ=⎝⎛θ110